Professional Documents
Culture Documents
Summary. Problems in structural dynamics lead quite often to stiff second order
ODEs. Solving them as first order ODE with explicit methods results in very small
stepsizes due to high frequency oscillations and instability. Implicit methods, on the
other hand, are not well-suited for real-time applications and very large problems.
Thus, there is a need for low order methods that allow relatively large steps and
that are computationally not expensive. In this paper, we introduce a class of RungeKutta-Nystrom methods that meet the requirements above. The methods generalize
the well-known Stormer's rule and allow a maximization of the stability domain.
Key words: Runge-Kutta-Nystrom methods, stability, second order ODE, stiff mechanical system, structural dynamics
140
C. Lunk, B. Simeon
+f
where a(u) stands for the stress tensor, e.g., a(u) = 1/2C ("\i'u + V'uT) in
case of Hooke's law, while f and p denote external forces and mass density.
If we pass to the weak form of the equations of motion and apply a RitzGalerkin method, e.g., a finite element method, we obtain a second order
system of ODEs in the unknown vector q(t) E IRn of node variables (discrete
displacements). These semi-discretized equations have the form
q(to) = Qo,
q'(to) = qb
(1)
with positiv definite mass matrix M, stiffness term k(y), damping matrix D
and external force vector b(q, t) (in obvious dimensions).
We remark that the assumption of small strains leads usually to a linear
stiffness term k(q) = Kq with stiffness matrix K while large strains introduce
some nonlinearities in k(q). Furthermore, Rayleigh damping is commonly used
in structural dynamics, i.e., D = fJ1M + fJ2K with scalars fJ1 and f32
y(to) =Yo,
y'(to) = Yb
(2)
141
Runge-Kutta-Nys trom Methods. In order to obtain larger stability domains, we focus on the class of Runge-Kutta-Nystro m (RKN) methods that
is especially adapted to second order ODEs
Y (to) = Yo,
y" = f(t,y,y'),
y' (to)
y~.
k~ = f(t
h2
j=l
j=l
L aijkj, y~ + h L aijkj)
Y1 =Yo+ hy~
L bikL
y(t 0
+ h)
and
1, ... , s;
h2
y~
i=l
y~
+ h 2:: bi k~.
(3)
i=l
( Y';+l )
Ym+l
= L ( Y';)
Ym
142
C. Lunk, B. Simeon
methods [HuOO] and that we adopt here. More precisely, we have weak
stability if
-( ..72 + 1) <
2
'7
< . 72 + 1
vl-
and
We skip the details of the coefficient search and refer the interested reader
to [Lu02]. Summarizing the results, it turns out that there exist methods
with a stability domain up to e = 2s. However, these methods are only explicit in the position variables and implicit in the velocities. They can be
viewed as generalizations of the well-known Stormer's rule. In order to keep
the computational effort for the implicit velocity terms low, we succeeded
in deriving particular coefficient sets that feature an SDIRK-like structure
with respect to the velocity coefficients. Table 1 gives some examples of the
methods. Moreover, Figs. 2-5 show the corresponding stability domains in the
(~,B)-plane. Due to the special structure of the second order system (2) and
the discretization (3), the notion of stability domains C <r is not applicable
here. We observe that the methods are unconditionally stable with respect to
the damping factor ~ while with respect to the frequency w, the stepsize is
limited by e :::; 2s ==? hmax = 2s jw. In case of Stormer's rule, however, we
have e::::: 2(s- 1) = 2.
10
12
12
10
10
86
2.
5.
3.
= 0.
= 2s .
143
:=
qm
4 Numerical Examples
Our test example is the 2D-pylon with time-dependent hanging load of
Fig. 1. Its segments are modelled by Euler-Bernoulli-beams, and the FEdescritization leads to a total of 107 nodes with 3 degrees of freedom each.
After elemination of boundary conditions we obtain a second order ODE with
315 unknowns. The highest and lowest eigenfrequencies range from about
~ 741.
2.649 105 to 357, i.e., Wmax
Wmin
Table 1. Different Runge-Kutta-Nystrom coefficients, p: method order,
of stages. In Table 1 'a., Stormer's rule is obtained for ry = 0.
Tab.l 'a.
Newmark
method.
Tab.l 'c.
RKN with
8 = 3,
p= 2.
c; 1 11
1
2-ry
,/ a;j 11
'"Y
2 ~
+--
b;
Tab.l 'b.
RKN with
8 = p = 2.
9 H 25
1 72 72
I24
5
97
360
I
45 1
97
Tab.l 'd.
RKN with
8 = p = 3.
20 100 25
8 :
number
11.!
2
I~ ~I
0
19
~7
33IT33
1_92 ~9
1 11
18 6
1
6
Table 2. Pylon: Relative error and CPU time (h:min:s), 2GHz Pentium IV.
model
linear
nondamped
linear
damped
nonlinear
nondamped
nonlinear
damped
I
I
I
ode15s
ode45
8.4. 10- 5
(0:02:02)
8.4. 10-5
(0:02:02)
8.4. 10- 5
(0:02:02)
2.8. 10- 6
(0:04:23)
8.6 10- 5
(0:02:49)
8.6. w- 5
(0:02:48)
5.0. 10- 3
(0:01:42)
2.5. w- 6
(0:05:06)
3.2 10- 2
(0:12:43)
1.9. 10-2
(0:12:55)
1.9 10- 2
(0:12:54)
1.5 . 10-5
(0:23:18)
1.9 10- 2
(0:11:00)
4.6. 10- 4
(0:11:15)
4.1. 10- 3
(0:06:43)
2.5. 10- 6
(0:20:56)
144
C. Lunk, B. Simeon
Table 2 compares our RKN-methods of Tab. 1 with the two MATLABsolvers ode45 (RK 5th order) and ode15s (BDF 1st and 2nd order). We
took the default settings for these codes and applied the RKN-method with
maximum stepsize. As the relative errors and computation times both for
a damped and undamped model as well as a linear and nonlinear stiffness
term show, the RKN-methods are competitive and efficient.
References
(HNW93] Hairer, E., N0rsett, S.P., Wanner, G.: Solving Ordinary Differential Equations I, Nonstiff Problems. 2. edition. Springer, Heidelberg (1993)
(HW96] Hairer, E., Wanner, G.: Solving Ordinary Differential Equations II, Stiff
and Differential-Algebraic Problems. Springer, Heidelberg (1996)
(HuOO]
Hughes: The Finite Element Method. Dover, Princeton (2000)
(Lu02]
Lunk, Ch.: Runge-Kutta-Nystrom-Verfahre n zur Losung steifer Differentialgleichungen aus der Strukturmechanik. MA Thesis, University of
Karlsruhe (2002)